1. Project Type and Tech Stack Language: Rust (2021 edition)
   Build System: Cargo (Cargo.toml / Cargo.lock)
   Binary: anamnesic (entry point: src/main.rs)
   Description: "Local coding agent — plan → act → verify. Fusion of TinyCoder + llm-on-legacy-gpus" Key Dependencies Category Crates TUI ratatui (0.29, with unstable-rendered-line-info), crossterm (0.28) Async runtime tokio (1, full features) LLM / HTTP reqwest (0.12), serde/serde_json, axum (0.8, with ws) Local inference ocl (optional, GPU), bytemuck, rand Database rusqlite (0.31, bundled) Terminal/PTY portable-pty (0.8) — ConPTY on Windows, PTY on Unix Search/Regex regex (1), ignore (0.4) Diff/Text diffy (0.4), pulldown-cmark (0.12), unicode-segmentation, unicode-width CLI clap (4, derive) Testing wiremock (0.6, dev-dependency)
2. Directory Structure Overview C:\Users\luann\Documents\GitHub\anamnesic-coder/ ├── Cargo.toml ├── Cargo.lock ├── src/ │ ├── main.rs # Entry point, CLI parsing, REPL, session resume │ ├── agent/ │ │ ├── mod.rs │ │ ├── agent_loop.rs # Core agent loop, tool execution, MCP integration, search_code │ │ ├── executor.rs # Step execution, search_code + grep_context helpers │ │ ├── planner.rs │ │ └── state.rs # AgentState (session, MCP clients, todos, transactions) │ ├── bench/ # Model benchmarking (local + cloud) │ ├── compressor/ # L1/L2 output compression │ ├── config/ # Settings, global settings, allow/deny lists │ ├── hw_recommend/ # Hardware detection + model recommendations │ ├── llm/ # LLM router, client, provider chain, GGUF inference, embedder │ │ └── infer/ # GGUF reading, tokenizer, engine, GPU kernels │ ├── mcp/ # MCP client (JSON-RPC stdio subprocess) │ ├── memory/ # Short-term + long-term (SQLite) memory │ ├── providers/ # Provider store, verification + models.dev catalog client │ ├── repo/ # Repo scanner, SymbolIndex, RepoMapGenerator, context │ ├── skills/ # Skills system (SKILL.md packs) │ ├── terminal/ # PTY + WebSocket server for browser TUI │ │ ├── pty.rs # TerminalSession (portable-pty wrapper) │ │ ├── session.rs # TerminalSessionManager │ │ ├── server.rs # Axum HTTP + WS server │ │ ├── shell.rs # Default shell detection (Windows/Unix) │ │ └── websocket.rs │ ├── tools/ # Agent tool implementations │ │ ├── fs.rs # FileTools (path-scoped, symlink, allow/deny) │ │ ├── transaction.rs # WorkspaceTransaction, diff, checksums │ │ ├── shell.rs # Shell command execution (cross-platform) │ │ ├── git.rs # Git tools │ │ ├── test.rs # Verification gates (cargo test, pytest, npm) │ │ ├── background.rs # Background task manager (C9) │ │ └── web.rs # Web search + fetch │ ├── types/ # Shared types (error, action, plan) │ └── ui/ # Ratatui TUI │ ├── mod.rs # Main TUI (~4023 lines) │ ├── file_search.rs # Fuzzy file search (Ctrl+P) │ ├── diff_render.rs │ ├── live_wrap.rs │ ├── pager_overlay.rs │ └── line_truncation.rs ├── docs/ │ ├── adr/ # 15 ADRs (0001–0015) │ └── gap-analysis-2026-08.md ├── prompts/ # System/coder/planner/fixer prompt files ├── sandbox/ # Dockerfile + runner.sh └── memory_data/memory.db # SQLite long-term memory
3. Existing Files Related to Backlog Items A. Symlink Handling Key files:

• src/tools/fs.rs — FileTools with workspace containment logic
• Lines 27–50: normalize_workspace_path() with Windows verbatim (\?) prefix support
• Line 134: Unix-only test rejects_symlink_escape_outside_workspace — creates a symlink to /etc and verifies the agent cannot read/write through it
• Lines 518+: file_type.is_symlink() check in list_files/list_tree traversal
• Lines 193–221: Windows-only tests verbatim_prefix_workspace_is_supported and absolute_verbatim_path_inside_workspace_resolves
• src/tools/transaction.rs — WorkspaceTransaction
• Line 275: if path.is_symlink() { continue; } — symlinks are skipped during workspace snapshot
• src/agent/agent_loop.rs
• Line 1642: if file_type.is_symlink() { continue; } in search_workspace_without_rg()
• docs/adr/0008-robustness-and-floating-point-safety.md — Documents path traversal, symlink escape, and injection safety fixes Windows-specific gap: Only the Unix symlink escape test exists (#[cfg(unix)]). There is no Windows junction/reparse-point symlink test. The Windows FileTools tests focus on verbatim path prefixes instead. B. Search/Lookup Functionality (Regex-based) Key files:
• src/agent/agent_loop.rs
• Lines 1622–1686: search_workspace_without_rg() — fallback regex search when rg is not installed. Uses regex::Regex::new(pattern) with literal fallback, scans files skipping symlinks, .git, target, node_modules, memory_data
• Lines 850: "search_code" tool dispatch
• Lines 1793–1794: Tool definition: "search_code" — "Search workspace text using a regex or literal pattern."
• src/agent/executor.rs
• Lines 302–370: search_code(state, pattern) function — prefers rg (ripgrep), falls back to search_workspace_without_rg()
• Lines 426–485: grep_context() helper
• src/repo/scanner.rs — RepoMapGenerator + SymbolIndex
• Lines 130–194: Regex-based symbol extraction for Rust (fn/struct/enum/trait), Python (def/class), JS/TS (function/class)
• Lines 197–241: Tests for repo map generation and symbol index search
• src/ui/file_search.rs — Fuzzy file search for Ctrl+P
• walk_files(), search_files(), fuzzy_match() (subsequence matching with scoring)
• 8 inline tests
• src/compressor/layer1.rs and layer2.rs — Regex-based output compression (timestamps, UUIDs, hashes, JWT, etc.)
• src/llm/router.rs — capability_lookup_reflects_the_active_provider_catalog() test C. MCP Server Management Key files:
• src/mcp/mod.rs — Complete MCP client implementation (235 lines)
• McpServerConfig struct: command, args, env
• McpClient::connect() — spawns subprocess, pipes stdin/stdout
• JSON-RPC 2.0 protocol: initialize, tools/list, tools/call
• list_tools() → converts to ToolDef format
• has_tool() / call_tool() for dispatching
• Tests: config creation/equality, fake_mcp_server_process (spawns self as fake MCP server via ANAMNESIC_FAKE_MCP_SERVER=1)
• src/agent/agent_loop.rs
• Lines 675–683: connect_mcp_clients() — connects all configured MCP servers at startup
• Lines 685–717: try_mcp_tool() — dispatches tool calls to MCP clients with approval policy
• Lines 1962–1964: MCP tools are enumerated and merged with native tools
• Lines 2406–2454: MCP tool tests (no clients, ask policy, allow policy, run policy)
• src/config/settings.rs
• Line 123: pub mcp_servers: Veccrate::mcp::McpServerConfig in Config
• src/agent/state.rs
• Line 40: pub mcp_clients: Veccrate::mcp::McpClient in AgentState
• docs/adr/0012-mcp-client.md — ADR for MCP stdio client D. Session Management Key files:
• src/agent/state.rs
• AgentState struct fields: session (ShortTermMemory), session_id (Option), session_persist, last_persisted_seq, last_persisted_summary
• ensure_session() — lazily creates persistent session record
• persist_session() — writes transcript to DB with compaction summary
• load_session_into_state(id) — restores session from DB, injects summary if present
• reset() — clears session, sets session_id = None for fresh start
• start_turn() — begins WorkspaceTransaction
• src/memory/short_term.rs — ShortTermMemory (in-context conversation history with token budget)
• src/memory/log.rs — LongTermMemory (SQLite-backed, session CRUD, vector store for memory_search)
• src/main.rs
• Lines 111–116: --resume / --cont CLI flags
• Lines 298–300: continue_latest_session() on startup
• Lines 441–460: resume_session() / continue_latest_session() — finds latest session for workspace and restores it E. Terminal/TUI Components Key files:
• src/ui/mod.rs — Main TUI (~4023 lines)
• Ratatui + Crossterm backend
• Focus modes: Sidebar, Messages, Input, Editor
• Slash commands: /info, /help, /status, /model, /provider, /reset, /resume, /continue
• Approval modal (a allow once, s allow session, d/Esc deny)
• File search overlay (Ctrl+P)
• src/terminal/ — Web-based terminal via WebSocket
• pty.rs: TerminalSession using portable-pty (ConPTY on Windows)
• session.rs: TerminalSessionManager (create/get/remove/resize sessions)
• server.rs: Axum HTTP + WS server, serves xterm.js SPA
• shell.rs: Platform-aware default shell detection
• src/main.rs
• Commands::Tui — launches ui::run_ui()
• Commands::Serve — launches terminal::server::serve() for browser access
• Auto-detects TUI mode when stdin/stdout are terminals

4. Key Entry Points and Main Modules Entry Point Location main() src/main.rs:259 run_agent_loop() src/agent/agent_loop.rs ui::run_ui() src/ui/mod.rs terminal::server::serve() src/terminal/server.rs repl() src/main.rs:384 build_router() src/main.rs:202
5. Testing Infrastructure Framework: Rust built-in #[test] (no external test framework like rstest or proptest)
   Test count: 298 test functions across the codebase
   Dev dependency: wiremock (0.6) — available but not heavily used in the current inline tests Test distribution by file: File src/agent/agent_loop.rs src/tools/fs.rs src/tools/shell.rs src/llm/router.rs src/memory/short_term.rs src/config/settings.rs src/compressor/layer1.rs src/ui/file_search.rs src/tools/transaction.rs src/agent/state.rs Test patterns:

• Most tests use temp directories (std::env::temp_dir()) with cleanup
• MCP tests use a clever self-spawning fake server (ANAMNESIC_FAKE_MCP_SERVER=1)
• Platform-gated tests: #[cfg(unix)] for symlink tests, #[cfg(windows)] for verbatim path tests
• Agent loop tests use a test_state() helper that creates isolated temp workspaces

6. Competitive Backlog References (C1, C2, C8) Primary reference: docs/adr/0015-competitive-backlog.md C1 — Path-scoped permissions ✅ Implemented

• Files: src/config/settings.rs, src/tools/fs.rs, src/tools/shell.rs
• path_allowlist, path_denylist, block_workspace_escape config
• FileTools::resolve() enforces workspace containment
• Symlink escape prevention (Unix tested, Windows not explicitly tested for junctions) C2 — Symbol / LSP search ✅ Implemented (regex-based, not full LSP)
• Files: src/repo/scanner.rs, src/agent/agent_loop.rs
• SymbolIndex::build() extracts symbols via regex (Rust, Python, JS/TS)
• SymbolIndex::search() and search_type()
• RepoMapGenerator::generate_map() produces compact repo map
• TODO.md notes: "Melhora futura: símbolos via LSP/tree-sitter (ver C2 no backlog competitivo)" — indicates the regex approach is a stepping stone to full LSP C8 — Notebook editing / timers / image gen ❌ Deferred
• ADR states: "Requires heavy dependencies (tree-sitter, external APIs). Not in scope."
• TODO.md: "Notebook editing (Claude/Cursor), timers/cron (Antigravity), image gen (Antigravity). Avaliar público-alvo." Other C-items (for context): Item Status C3 Parallel sub-agents ✅ C4 Skills system ✅ C5 Checksums/change-tracking ✅ C6 Extended thinking ✅ C7 Adversarial verification ✅ C9 Background tasks ✅

7. Notable Gaps / Observations
8. Windows symlink testing is absent. The only symlink escape test is #[cfg(unix)]. On Windows, NTFS junctions/reparse points are not tested for escape scenarios.
9. Search is regex + ripgrep fallback, not LSP. search_code uses rg or a manual directory walk with regex::Regex. C2 was "implemented" as regex symbol extraction, but the TODO explicitly flags LSP/tree-sitter as the future direction.
10. No integration test harness with mock LLM providers. The dev-dependencies include wiremock but the inline tests use a self-spawning MCP fake server pattern instead. There is no mock Ollama/cloud provider for end-to-end agent loop testing.
11. Testing is all inline #[test] modules. No separate tests/ directory or integration test binaries (except the self-spawning MCP server trick).
12. C8 is explicitly deferred — notebook editing, timers, and image generation are out of scope per ADR 0015.